Dense and Reliable Dsm Generation from Vhr Stereo Pairs in Urban Environments

Photogrammetric terrain reconstruction from aerial and space stereopairs of images occupies a prominent place in cartography and remote sensing. Stereo vision systems determine depth from two or more images using automated techniques. The most important and time consuming task for a stereo vision system is the registration of both images, i.e. the matching of corresponding pixels. Area based stereo attempts to determine the correspondence for every pixel, which results in a dense depth map. Correlation is the basic method used to find corresponding pixels. However, correlation assumes that the depth is equal for all pixels of a correlation window, which is violated at depth discontinuities. The result is that object borders are blurred and small details or objects are removed, depending on the size of the correlation window. In this paper, we focus on the generation of reliable surface models using dense techniques. An overview is given of correlation-based techniques using adaptive windows. These adaptive techniques separate forefrom background information in a correlation window using structure specific information (e.g. gradient, segmentation) and are able to compensate for the blurring effect that occurs at object boundaries. The result is a dense surface model with an emphasis on reliability. A case study that compares the different techniques is presented.